**Breaking Down the Barriers in Operations Research with OptiMind**
I’m super excited to share with you the latest innovation from Microsoft Research – OptiMind, an AI-based system that can convert natural language descriptions of complex optimization problems into mathematical formulations that optimization solvers can execute. This is a game-changer in operations research, where translating business intent into mixed-integer linear programs typically requires expert modelers and days of labor.
**What’s the big deal about OptiMind?**
OptiMind is a specialized 20B parameter model in the gpt oss transformer family, with around 3.6B parameters active per token. It takes natural language descriptions of optimization problems as input and outputs a mathematical formulation, including executable Python code that uses GurobiPy. The generated script defines decision variables, constraints, and objectives, calls the Gurobi solver, and prints the optimal objective value and solutions.
**How was OptiMind developed?**
The model is based on openai/gpt-oss-20b, fine-tuned using cleaned optimization datasets such as OR-Instruct and OptMATH. The research team used a combination of optimization expertise and LLM training to fine-tune the model. They classified problems from the datasets into 53 seed classes, such as set cover, flow shop scheduling, or traveling salesman problem, and then ran the gpt-oss-20b-base model on a sample of problems. They selected cases where the model output disagreed with the ground truth, and optimization specialists examined these cases to identify recurring formulation errors.
**How does OptiMind work at inference time?**
At inference time, OptiMind behaves as a multi-stage system. First, it classifies each test instance into one of the 53 optimization classes used during error analysis. Then, it augments the prompt with the error abstract and trace pairs related to that class. The model generates a reasoning hint, the mathematical formulation, and the GurobiPy code. When more compute is available, the system can apply self-consistency with majority voting, generating multiple candidate scripts, executing them, and selecting the answer that appears most frequently within set numerical tolerances.
**How well does OptiMind perform?**
On cleaned versions of IndustryOR, Mamo-Complicated, and OptMATH, OptiMind significantly improves answer accuracy. The fine-tuned model improves formulation accuracy by 20.7% across multiple optimization benchmarks, with further gains when test-time scaling methods such as self-consistency and multi-round feedback are applied. OptiMind outperforms other open-source models of comparable or larger size, reaching performance competitive with proprietary frontiers models such as GPT-o4 mini and GPT-5 under the evaluation settings.
**Key Takeaways**
1. OptiMind is a 20B parameter model that takes natural language optimization problems as input and outputs both a mathematical formulation and executable GurobiPy code.
2. The model is fine-tuned on cleaned optimization datasets and evaluated on expert-validated benchmarks.
3. OptiMind uses class-based error analysis and expert-written hints for 53 optimization classes, then applies these hints both in data cleaning and at inference time.
4. The framework improves formulation accuracy by 20.7% across multiple optimization benchmarks compared to the base model.
5. OptiMind-SFT is released as microsoft/OptiMind-SFT on Hugging Face and as microsoft-optimind-sft in Azure AI Foundry, where it can be served through SGLang as an OpenAI-compatible endpoint.
**Try the Model Weights and Technical details**
Head over to Hugging Face to access the model weights and technical details. You can also try out OptiMind-SFT in Azure AI Foundry, where it can be served through SGLang as an OpenAI-compatible endpoint, enabling seamless integration into decision support pipelines for supply chains, manufacturing, logistics, and scheduling.
